Kinetic equilibration rates for granular media and related equations: entropy dissipation and mass
45
pages
English
Documents
Obtenez un accès à la bibliothèque pour le consulter en ligne En savoir plus
Découvre YouScribe en t'inscrivant gratuitement
Découvre YouScribe en t'inscrivant gratuitement
45
pages
English
Documents
Obtenez un accès à la bibliothèque pour le consulter en ligne En savoir plus
Kinetic equilibration rates for granular media and related equations: entropy dissipation and mass transportation estimates Jose A. Carrillo?, Robert J. McCann†, Cedric Villani‡ September 14, 2004 Abstract The long-time asymptotics of certain nonlinear, nonlocal, diffusive equations with a gradient flow structure are analyzed. In particular, a result of Benedetto, Caglioti, Carrillo and Pulvirenti [4] guaranteeing eventual relaxation to equilib- rium velocities in a spatially homogeneous model of granular flow is extended and quantified by computing explicit relaxation rates. Our arguments rely on estab- lishing generalizations of logarithmic Sobolev inequalities and mass transportation inequalities, via either the Bakry-Emery method or the abstract approach of Otto and Villani [28]. Mathematics Subject Classification: 35B40, 35K55, 35K65, 35Q72. Keywords: rates of convergence, generalized log-Sobolev inequalities, Wasserstein distance, inelastic collision models. Contents 1 Introduction 2 2 Main results 7 3 Preliminary computations 15 3.1 Second variation of entropy F (?) under displacement . . . . . . . . . . . 15 3.2 Dissipation of entropy dissipation . . . . . . . . . . . . . . . . . . . . . . 17 ?Departamento de Matematica Aplicada, Universidad de Granada, 18071 Granada, SPAIN. †Department of Mathematics, University of Toronto, Toronto, Ontario, M5S 3G3, CANADA.
Voir
- entropy functional
- sobolev inequalities
- d? ≥
- dissipation functional
- euler-lagrange equation
- dissipation method
- diffusion
- behaviour along line
Kinetic equilibration rates for granular media and related equations: entropy dissipation and mass transportation estimates
Jose A. Carrillo, Robert J. McCann†, Cedric Villani‡§
September 14, 2004
Abstract
The long-time asymptotics of certain nonlinear, nonlocal, di usiv e equations withagradientowstructureareanalyzed.Inparticular,aresultofBenedetto, Caglioti, Carrillo and Pulvirenti [4] guaranteeing eventual relaxation to equilib-rium velocities in a spatially homogeneous model of granular
o w is extended and quantiedbycomputingexplicitrelaxationrates.Ourargumentsrelyonestab-lishing generalizations of logarithmic Sobolev inequalities and mass transportation inequalities, via either the Bakry-Emery method or the abstract approach of Otto and Villani [28].
Mathematics Subject Classication:35B40, 35K55, 35K65, 35Q72. Keywords:of convergence, generalized log-Sobolev inequalities, Wassersteinrates distance, inelastic collision models.
Contents
1 Introduction
2 Main results
3 Preliminary computations 3.1 Second variation of entropyF( . . . . . . . . . . .) under displacement 3.2 Dissipation of entropy dissipation . . . . . . . . . . . . . . . . . . . . . .
DepartamentodeMatematicaAplicada,UniversidaddeGranada,18071Granada,SPAIN. †Department of Mathematics, University of Toronto, Toronto, Ontario, M5S 3G3, CANADA. ‡3946F,6-eCedyLnoFRANx07,CE.UME,APelocmroNSelaeupeuridereonLy § mccann@math.toronto.edu, cvillani@umpa.ens-lyon.fre-mails: carrillo@ugr.es,
1
2
7
15 15 17
4 The proofs 4.1Uniformlyconvexconnement........................ 4.2Uniformlyconvexinteraction,xedcenterofmass............. 4.3 Degenerately convex interaction, perturbative argument . . . . . . . . . . 4.4Degeneratelyconvexinteractionwithdi usion............... 4.5 Treatment of moving center of mass . . . . . . . . . . . . . . . . . . . . . 5 Rates of convergence inL1
A The Cauchy problem and smooth approximations A.1 Linear di usion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.2 Nonlinear di usion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1 Introduction
18 19 22 23 26 28
30
32 33 38
This paper is devoted to the asymptotic behavior of solutions of the equation ∂ ∂t=r [r(U0() +V+W)],(1.1) where the unknown(t,) is a time-dependent probability measure onRd(d1),U: R+→Ris a density of internal energy,V:Rd→Ris a connemen t potential and W:Rd→Ris an interaction potential. The symbolrdenotes the gradient operator and will always be applied to functions, whilerstands for the divergence operator, and willalwaysbeappliedtovectorelds(orvectorvaluedmeasures).Inthesequel,we identify both the probability measure(t,) =twith its densitydt/dxwith respect to Lebesgue, and thus, we use the notationdt=d(t, x) =(t, x)dx shall make. We precise convexity assumptions aboutU,V,Wlater on; for the moment we just mention that it is convenient to requireWto be symmetric (∀z∈Rd,W( z) =W(z)), andU to satisfy the following dilation condition, introduced in McCann [25]: 7 →dU( d) is convex nonincreasing onR+.(1.2)
The most important case of application isU(s) =slogsnternalwh,ihcitnedseiieht energywithBoltzmann’sentropy,andyieldsalineardi usionterm,, in the right-hand side of (1.1). Equations such as (1.1) appear in various contexts; our interest for them arose from their use in the modelling of granular
o ws: see the works of Benedetto, Caglioti, Car-rillo, Pulvirenti, Toscani [3, 4, 31] and the references there for physical background and mathematical study (a short mathematical review is provided in [34, chapter 5]). Let us just recall the most important facts. To equation (1.1) is associated an entropy, orfree energy: F() =ZRdU()dxZRd(x21+)ZRdRdW(x y)d(x)d(y).(1.3) +V(x)d
2
This functional can be split into the sum of an internal energyU, a potential energyV and an interaction energyWthe three terms on the right-, corresponding respectively to hand side of (1.3). A simple computation shows that, at least for classical solutions, the time-derivative ofF() along solutions of (1.1) is the negative of D()ZRd||2d,(1.4)
where
r[U0() +V+W].
(1.5)
The functionalDwill henceforth be referred to as theentropy dissipation functional. SinceDis obviously nonnegative, the free energyFacts as aLyapunov functionalfor equation (1.1). In many cases of interest, the competition betweenU,VandWdetermines a unique minimizer∞forF, as shown in [25]. In this paper, our conditions onU,V,Wwill ensure that this is indeed true — except in certain situations where the minimizer will only be unique up to translation. A natural question is of course to determine whether solutions to (1.1) do converge to this minimizer ast→+∞, and how fast. To formulate this problemmoreprecisely,onerstneedstodecidehowtomeasurethedistancebetween and∞ this paper this . Inwill usually be achieved by the relative free energy, improperly calledrelative entropy: F(|∞) =F() F(∞).(1.6) Thus we intend to prove thatF((t,)|∞) converges to 0 ast→+∞, and estimate the speed of convergence. Let us mention here one of our main results, and relate it to previous work. A few years ago, Benedetto, Caglioti, Carrillo and Pulvirenti [4] studied equation (1.1) in the case (arising in the modelling of granular material) whenU(s) = slogs,V(x) =|x|2/2, W(z) =|z|3,d= 1 ( >, 0). Via the study of the free energyF, they proved convergence to equilibrium in large time, without obtaining any rate; here we shall prove exponential convergence at an explicit rate. Moreover, our result holds for any dimension of space, for interaction potentials which are perturbations of|z|3(while the method in [4] needsW(z) =|z|3heavily andd= 1 in order thatFbe a convex functional), and we shall also prove exponential convergence when= 0. Our proofs are based on the so-calledentropy dissipation method, which consists inbounding below the entropy dissipation functional(1.4)in terms of the relative en-tropy rst(1.6). At this is a quite technical task in view of the complexity of sight functionals (1.4) and (1.6). Furthermore, the value ofF(∞) is not explicitly known, since the Euler-Lagrange equation for the minimizer of (1.3) seems to be unsolvable — thus our strategy may seem doomed from the very beginning. But the particular struc-ture of equation (1.1) will allow the use of powerful methods taking their roots in the theory of logarithmic Sobolev inequalities.
3
At this point we may recall one of the most fundamental results in this theory, due to Bakry and Emery [2]. Consider the case whenU(s) =slogs,W= 0, and assume that Vis uniformly convex, in the sense that there exists >0 such that D2V , I(1.7) whereIis the identity matrix onRd, and the inequality holds in the sense of symmetric matrices. Then D()2 F(|∞) (1.8) Assuming without loss of generality thatRe V= 1, this can be rewritten as ZRd|r(log+V)|2d2ZRd(log+V)d.(1.9)
This is one of the many forms taken bylogarithmic Sobolev inequalities task we. The undertake here is to generalize the functional inequality (1.8) to handle the nonlocal nonlinearity introduced by an interaction potentialW to now, the most noticeable. Up generalization of (1.8) had been the replacement of the Boltzmann entropyU(s) =slogs by other functionals associated with nonlinear but still local di usion, as in works of Carrillo, Jungel, Markowich, Toscani, Unterreiter, Dolbeault, del Pino and Otto [12, 9, 18, 27]. There are at least two general methods to prove inequalities such as (1.9) (and only two, so far as we know, which are robust enough to be used in our context); we shall work out both of them. The rst one, inherited from the seminal work of Bakry and Emery, goes via the study ofthe second derivative inof the relative entropy functional. Indeed, many cases of interest, a direct comparison of the entropy with its dissipation is a very dicult task, buta comparison of the entropy dissipation with the dissipation of entropy dissipation is much easier. The surprising success of this method [1, 11, 12, 9] (see [23] for a tentative user-friendly review on these techniques), which may seem hard to believe at rst, can be explained at a heuristical level by the conceptual work of Otto [27], who showed that the relation of the equation (1.1) to the free energy (1.3) has the structure ofagradientow.Sinceitiswell-knownthattheasymptoticbehaviorofthetrajectories ofagradientowarecloselylinkedtotheconvexitypropertiesofthecorresponding functional, this suggests that di eren tiating twice is a natural thing to do. In fact, the relevant notion of convexity, in this context, is thedisplacement convexityintroduced by McCann [25], whose denition will be recalled below. In Section 3 of Otto and Villani [28] it is proven in an abstract framework that, at least from the formal point of view,uniformdisplacement convexity implies an inequality of the same type as (1.8). The illumination provided by this point of view is suggested brieybythefollowingexample.Assumeaconvexfunctionf:R →[0,∞) attains its minimum valuef(w) = 0 atw From uniform convexity,= 0.f00(w), it is easy to deduce the inequalities f(w) 2(gdra1f)2(w)0,(1.10)
4
